CELLO
收藏资源简介:
CELLO数据集包含523条英语指令,这些指令均来源于真实场景中人工整理的数据。其主要目标是衡量大型语言模型(LLMs)对复杂指令的理解能力。评估内容涵盖十个子任务,涉及复杂任务描述和复杂输入相关方面。评估方法采用代码评估。
The CELLO dataset encompasses 523 English instructions, meticulously curated from real-world scenarios. Its primary objective is to evaluate the ability of Large Language Models (LLMs) to understand complex instructions. The evaluation encompasses ten subtasks, covering aspects of complex task descriptions and complex inputs. The assessment method employed is code-based evaluation.
CELLO数据集概述
数据集简介
- CELLO是一个用于系统评估大语言模型复杂指令理解能力的基准测试工具(AAAI 2024)。
- 主要特点:
- 设计了8个复杂指令特征
- 从真实场景构建综合评估数据集
- 建立了4个评估标准并开发了相应指标
- 比较了中英文导向模型在复杂指令理解上的表现
数据内容
- 数据位置:data/
- 所有样本均已匿名化处理
评估系统
- 评估脚本:
eval.sh - 模型实现位置:code/evaluators/
- 评估结果位置:results/
评分系统
- 评分脚本:
score.sh - 评分器实现位置:code/scorers/
- 评分结果位置:scores/
引用格式
bibtex @inproceedings{he2024can, title={Can Large Language Models Understand Real-World Complex Instructions?}, author={He, Qianyu and Zeng, Jie and Huang, Wenhao and Chen, Lina and Xiao, Jin and He, Qianxi and Zhou, Xunzhe and Liang, Jiaqing and Xiao, Yanghua}, booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, volume={38}, number={16}, pages={18188--18196}, year={2024} }




